Analog implementation of linear transforms
Analog phase-shift elements connect each of a plurality of input nodes to each of a plurality of output nodes, wherein each component is adapted to produce a phase shift in a periodic signal processed therethrough. A linear transformation of a data set of discrete values of a given function provided as a set of analog signals to the input nodes is achieved by judiciously adjusting the signal amplitude produced at the output of the phase-shift components and summing the resulting output signals as required to simulate the transformation of interest.
1. Field of the Invention
This invention is related in general to devices that perform linear transformations, such as Fourier and wavelet transforms. In particular, the invention relates to a programmable analog circuit capable of performing any discrete linear transformation and its inverse.
2. Description of the Related Art
Linear transforms are very useful tools in science and technology. For example, Fourier transforms and inverse transforms are commonly utilized in many fields to analyze, design and implement signal processing, such as for defense and aerospace applications (radar, sonar, synthetic aperture radar, electronic warfare), medical diagnostic imaging (ultrasound, computed tomography, magnetic resonance imaging), telecommunications (broadband, wireless, digital video broadcasting), instrumentation and measurement (spectrum analysis, radio astronomy), and industrial vision (non-destructive testing, pattern recognition).
In all such applications, Fourier transforms are implemented with digital computers that utilize well known Discrete Fourier Transform (DFT) algorithms. For example, using current state-of-the-art processors, the algorithm known as the Fast Fourier Transform (FFT) can produce a 1024-point transform in approximately 10 μs, with a corresponding maximum data throughput of about 100 kHz. Multiple, staggered processors may be used operating in parallel in order to increase speed, but with increased power requirements that can exceed the capability of conventional bus configurations. Accordingly, the current implementation of DFT algorithms in digital computers is limited by hardware constraints.
Progress has been made in an effort to increase the processing speed of FFT algorithms by reducing the number of computations. For example, U.S. Pat. No. 5,987,005 describes an approach whereby DFT and inverse DFT operations are computed using the same computing device, thereby optimizing computational efficiency. The very high number of binary operations required to implement these algorithms and the increased power requirements associated with faster implementations remain a severe limiting factor in efforts to provide significant improvements in the speed of transform and inverse transform computation.
Therefore, there is still a need for a faster approach to the computation of linear transforms that does not also require unacceptably high power consumption. This invention has achieved this objective by effecting the transform and inverse-transform computations using analog devices.
BRIEF SUMMARY OF THE INVENTIONThe invention was derived while searching for an improved approach to perform Fourier transforms and inverse Fourier transforms. Accordingly, this disclosure is based primarily on an analysis of the classical representation of discrete Fourier transforms even though it was later realized that the solution provided by the invention is equally applicable to any other linear transformation. As such, the scope of the invention is not intended to be limited to Fourier transform applications.
As well understood in the art, the algorithm implemented in digital computers to calculate the discrete Fourier transform, Fm=F(km), of a function fn=f(xn) is defined by the following equations,
wherein the indices n and m are used to refer to discrete values of x and k, respectively, and the √{square root over (N)} factor (or an equivalent factor) is included to normalize the transform and inverse transform and to provide symmetry, as is well understood in the art. The imaginary exponential is defined by Euler's equation,
eiφ=cos(φ)+i sin(φ) (3)
where φ=±2πnm/N.
In searching for an analog implementation of this DFT algorithm, it was realized that its exponential components may be modeled effectively by the phase change of an electronic wave propagating through a conductive medium (such as a waveguide).
It is known that the velocity of propagation and attenuation of signals along a waveguide are functions of the geometry and materials used in its construction. That is, once the applied signal reaches a given point on the waveguide, the signal remains unaltered but for some phase and amplitude changes related to the distance of propagation. For example, referring to a waveguide of length L, as illustrated in
V0(t)=V0eiωt. (4)
where V0 is the maximum amplitude of the input signal; ω=2πf, f being the frequency of oscillation; and t is time. The output signal at the distance L along the waveguide thus can be represented by the equation
VL(t)=V0(t)A(L)exp(−iωL/v), (5)
where v is the velocity of propagation and A(L) is a parameter used to indicate the signal attenuation resulting from waveguide losses.
Equation 5 shows that any desired phase change can be obtained through a waveguide by judiciously selecting its length L. For example, as illustrated in
Vn(t)=V0(t)Anexp(−iωLn/v), (6)
where An and Ln are the signal attenuation and waveguide length, respectively, corresponding to output node n. Applying the principle of wave superposition to the illustrations of
where 1≦n≦N.
If the length of a waveguide, Ln, is selected such that
Ln=(2πnv)/(ωN), (8)
then Equation 7 becomes
Vn(t) can be expressed in conventional manner as a function of its maximum amplitude Vn as follows,
Vn(t)=Vneiωt. (10)
Combining Equations 4 and 10 with Equation 9 yields
which simplifies to
This equation illustrates that the input signal at point 0 in
The present invention is based on the recognition that Equation 11 has the same form as the mathematical representation of a discrete Fourier transform (Equation 1) and that, as such, it can be used advantageously to model the transform in an analog implementation. Carrying the concept forward to a total of M input nodes, each input node m may be connected to N output nodes by specific lengths of waveguides, as illustrated in
the same analysis used above with respect to a single output node yields the general relation
where φm is the phase of the mth input signal. This equation has the identical form of a Discrete Fourier Transform (DFT).
As one skilled in the art would readily recognize, phase changes that are integral multiples of 2π produce equivalent output signals (that is, phase changes of 2 kπ, where k is any integer, are equivalent). Therefore, Equation 12 can be written without loss of generality as
where mod(x,y) is defined as the remainder of the integral division of x by y. Using this idea of equivalent phase changes to compute the Inverse Discrete Fourier Transform (IDFT), the corresponding waveguide lengths can also be computed with the following, which provides the equivalent of the complex conjugate of the imaginary exponential:
An analysis of the speed of operation of such a waveguide implementation of a linear transform indicates an improvement of several orders of magnitude with respect to digital devices. Rearranging Equations 14 and 15 to determine the time required for the phase-shifted signals to reach the output, the maximum time is given by the relation Δt=2π/ω=1/f, f being the frequency of oscillation of the sine wave source. Using a radio frequency of about 10 MHz, which is moderate for modern signal generators, the computation time for a complete DFT or IDFT would be about 100 ns, which is two orders of magnitude faster than achievable with current state-of-the-art FFT processors. Using operating frequencies currently available for high-speed microprocessors (about one GHz), the computation time would be reduced to 1 ns, two additional orders of magnitude improvement. Furthermore, because of the parallel operation of the system, the computational speed is the same for any number N of input and output nodes.
The power requirements of the invention have been found to be similarly advantageous over prior-art digital implementations. In order to express Equation 5 with the waveguide configuration of
The relation expressed in Equation 11 with reference to the analog circuit of
Therefore, in most general terms, the invention consists of analog phase-shift components connecting each of a plurality of input nodes to each of a plurality of output nodes, wherein each component is adapted to produce a phase shift in a periodic signal processed therethrough. A linear transformation of a data set of discrete values of a given function provided as a set of signals to the input nodes is achieved by judiciously adjusting the signal amplitude produced at the output of the phase-shift components and summing the resulting output signals as required to simulate the transformation of interest. As a special case of the invention, those skilled in the art will recognize that real transforms can be implemented simply with two lengths of waveguide providing π and 2π phase shifts, respectively, and common phase inputs (constant φm).
It is noted that the purpose of computing a Fourier transform can be simply one of analysis, such as for determining the frequencies that comprise a signal. However, more general uses of the transformed data exist. For example, in order to filter all but a few frequencies of a signal, it is a simple matter to perform a DFT on the signal and then an IDFT of only the desired frequencies. This type of operation provides an ideal filter that can eliminate noise or unwanted information from a signal and the present invention, because of its flexibility of implementation, is advantageously suited for this type of application.
Referring to
It is also noted that the implementation of the invention disclosed above incorporates independent waveguides between each input and output nodes. Therefore, any structure capable of providing the phase-shift effect of a waveguide, such as a planar geometry, could be used in equivalent manner to implement a DFT. In addition, the phase-delay network can be implemented using either passive or active components. Programmable impedance devices may also be used to implement, alter, or tune the input, Stage-1, Stage-2, or Stage-3 sections.
Various other purposes and advantages of the invention will become clear from its description in the specification that follows and from the novel features particularly pointed out in the appended claims. Therefore, to the accomplishment of the objectives described above, this invention consists of the features hereinafter illustrated in the drawings, fully described in the detailed description of the preferred embodiment and particularly pointed out in the claims. However, such drawings and description disclose but one of the various ways in which the invention may be practiced.
The preferred embodiment of the invention is based on the general concept outlined above and the further realization that, inasmuch as phase changes of integral multiples of 2π are not detectable, the domain of phase changes Δφ occurring in a signal traveling through a waveguide are given by the relation
where mod(x,y) refers to the remainder of the integral division of x by y. This equation describes N equally spaced phase changes of increment 2π/N, which can be implemented in a network simply by using P (=kN, where k is an integer) phase-shift elements as illustrated in FIG. 6. Such a circuit may be tapped at any point 12 between fractional phase-shift elements 10 to acquire the desired shifted signal. It is clear that this serial circuit is equivalent to the waveguide configuration of FIG. 2. Each output of waveguide in
Thus, the multi-input-node linear transformation circuit of
F(t)=∫Ateiφ(ω)t∫f(ω)eiωtdωdt, (17)
where the internal integral is simply the Fourier transform of the original thermal noise.
If the amplification, A, is set to reduce to unity after a predetermined peak signal is attained, Equation 17 becomes
F(t)=∫f(ω)∫eiφ(ω)teiωtdtdω. (18)
As one skilled in the art would readily appreciate, as a result of the orthogonal property of imaginary exponential functions, ω=φ(ω), which means that the two imaginary exponentials are in phase. In turn, this means that φ=2 nπ. Therefore, simply by using the output of the series phase delay network as positive feedback to reduce to a unity-gain amplifier, the proper oscillation for the network of the invention is automatically created. Thus, applying this concept to the fractional phase-shift circuit of
As illustrated in
Based on the array output configuration illustrated in
If the amplitude-management amplifier 14 (see
where Rm,n,p is the signal amplitude-modification resistor 24 tapped into the pth phase-shift element 10 of the mth phase-shift circuit 16 to contribute to the nth output signal, On, at a corresponding nth output node 22; and Rg,n is the gain resistor 26 for the corresponding nth element of the transform vector of points. If Rm,n,p is selected such that
where δ(x=0)=1, δ(x≠0)=0, K is a reasonable resistance value, and γm is a constant; and if Rg,n is set equal to K/N, the output On then becomes
which is readily identified as the discrete Fourier transform of the N-element set
Anexp[i(φm−2πγm/N)]. (22)
Since all indices, including p, are limited to integral values, γm is similarly restricted. Moreover, phase shifts of multiples of 2π are undetectable and, therefore, can be discarded. Therefore, a programmable value for the required resistance may be obtained by the formula
As discussed above, it is possible to obtain phase-shifted signals from an N-element phase-shift oscillator. However, Equation 1 indicates that it is also necessary to control the complex-valued signal amplitude in order to implement the linear transform device of the invention. Equation 22 indicates that γm and φm are available to encode the phase component of the complex input, γm being simply an offset along the linear chain of phase-shift elements. However, use of γm to encode the input complex number would require runtime selection of the access points in the phase-shift network, which may be impractical. Therefore, a choice of γm=0 and the use of a “programmable” phase shift, φm at each input node may be adopted for implementation of the invention's input encoding scheme, as illustrated in the phase-management circuit 36 of
The real-valued amplitude at each input node is not so readily encodable as γm. Since a phase-shift oscillator's signal amplitude is only limited by the available voltage supply, the amplifier circuit must be designed to limit the gain of the entire circuit, including phase-shift elements, to unity. The amplitude at which this condition is satisfied may be manipulated by the circuitry. In essence, the amplifier circuit must adjust the overall gain to unity at the value of amplitude desired. Such a design can be realized with a peak detector and comparator adjusting the amplifier gain.
The phase-shift oscillator implementation reflected by
The concept of applying the desired 2π phase shift of the combined fractional network as the phase delay of a phase-shift oscillator may be extended in a manner that provides increased accuracy of the device. Since the inverse transform of a transform results in the original signal, the circuitry could be arranged to use a coupled transform/inverse transform construction as the feedback signal to a set of oscillation amplifiers. This sort of arrangement would help to dampen out the effects of component variations and essentially “lock in” a solution guaranteed to be a transform/inverse transform pair.
The circuit of
Although the analysis given above relates specifically to DFTs, it should be clear to those skilled in the art that the device of the invention is not restricted to performing such calculations. Equations 1 and 2 shows that any well-behaved function can be expressed in terms of Fourier coefficients. This being the case, any set of functions that can be used to model a discrete set of data can be expressed in the circuit of the invention through selection of the appropriate phase-shift tap (or waveguide lengths) and adjustment of the termination circuits. For example, the same circuit can be used in the computation of a wavelet transform.
It is clear that Equation 19 can be rewritten as
where
Equation 24 is the general definition of a linear transformation and, therefore, may represent any such transformation by simply setting the proper values of Rg,n and Rm,n,p. It is also clear that Equation 25, restricted to values of
where k is an integer, will construct a real-numbered transform. For negative coefficients, odd values of k are used. Since a discrete wavelet transform is, ultimately, just a real-valued transform, one need only set the resistance values to match the wavelet matrix.
Accordingly, without loss of generality, the use of DAUB04 was investigated as the “mother” wavelet function in a 16-element space, i.e., N=16. In this space, the transformation matrix is as reported in FIG. 15. If a fixed value is chosen for Rg,n=G, then the value for the signal amplitude modification resistors 24 (see
As would be clear to one skilled in the art, any other linear transformation can be achieved in the same manner using Equation 25 and the transformation matrix elements to determine the appropriate resistance values. Note that Equation 25 provides discrete phase values to match to the matrix elements of interest. In the event the matrix elements do not correspond to these discrete values, the signal generator array of the invention can be altered to provide more appropriate phase values either through the use of more phase-shift elements for a finer selection or through unequal phase-shift elements that provide the exact values required.
Various changes in the details, steps and components that have been described may be made by those skilled in the art within the principles and scope of the invention herein illustrated and defined in the appended claims. For example, while the serial circuit 16 of the invention has been described in terms of P preferably equal phase-shift elements 10, wherein each element produces a phase shift equal to 2π/P, it is clear that the invention could be practiced equivalently by serially connecting unequal phase-shift element so long as the cumulative phase shift remains set at 2π. Similarly, it is understood that analog phase-shift elements 10 are well known in the art and any kind can be used to practice the invention. A simple implementation of such an analog device is illustrated in FIG. 16.
It is noted that the invention has been described in terms of a phase-shift circuit 16 that produces a total shift of 2π radians using a series of P phase-shift elements 10. However, it is known in the art that a phase shift of π radians may be provided through the inverting input of an operational amplifier, particularly a fully differential amplifier. Therefore, with reference to
Finally, since real-valued transforms can be fully achieved with phase shifts of π radians, such a transform circuit could be implemented totally with differential amplifiers.
Therefore, while the present invention has been shown and described herein in what is believed to be the most practical and preferred embodiments, it is recognized that departures can be made therefrom within the scope of the invention, which is not to be limited to the details disclosed herein but is to be accorded the full scope of the claims so as to embrace any and all equivalent apparatus and procedures.
Claims
1. An analog device for effecting a linear transformation of input data corresponding to respective discrete values of a function, comprising:
- a plurality of input nodes for receiving analog signals representative of said input data;
- a plurality of analog phase-shift components connected to each of said plurality of input nodes for producing corresponding phase shifts in said analog signals; and
- a plurality of output nodes connected to said phase-shift components for producing analog output signals representative of output data corresponding to a linear transform of said discrete values of the function;
- wherein each of said phase-shift components includes a waveguide selected with a length suitable for producing a predetermined phase shift.
2. The device of claim 1, further including a plurality of amplitude-control elements corresponding to each of said output nodes to adjust respective outputs thereof to desired transform amplitudes.
3. The device of claim 2, wherein each of said amplitude-control elements includes a programmable resistor.
4. The device of claim 1, further comprising a computation device operating on said analog output signals to perform a predetermined operation, and an inverse-transform device to convert output data of said computation device to a set of corresponding processed values of said function.
5. The device of claim 4, wherein said inverse-transform device includes:
- a plurality of inverse-transform input nodes for receiving analog signals representative of said output data of the computation device;
- a plurality of analog inverse-transform phase-shift components connected to each of said plurality of inverse-transform input nodes for producing corresponding phase shifts in said analog signals representative of the output data of the computation device; and
- a plurality of inverse-transform output nodes connected to said inverse-transform phase-shift components for producing analog inverse-transform output signals representative of said processed values of the function.
6. The device of claim 5, further including:
- a plurality of amplitude-control elements corresponding to each of said output nodes to adjust respective outputs thereof to desired transform amplitudes; and
- a plurality of inverse-transform amplitude-control elements corresponding to each of said inverse-transform output nodes to adjust respective inverse-transform outputs thereof to desired inverse-transform amplitudes.
7. The device of claim 5, wherein each of said inverse-transform phase-shift components includes a waveguide selected with a length suitable for producing a predetermined phase shift.
8. The device of claim 5, wherein each of said inverse-transform phase-shift components includes at least one phase-shift element of a serially connected multiplicity of phase-shift elements, wherein said multiplicity of phase-shift elements produces a total phase shift equal to 2 kπ radians, k being an integer.
9. The device of claim 4, wherein said computation device performs a filtering operation.
10. The device of claim 4, wherein said computation device performs a convolution operation.
11. The device of claim 4, wherein said computation device performs a correlation operation.
12. The device of claim 4, wherein said computation device performs a linear-prediction operation.
13. The device of claim 4, wherein said computation device performs a matrix-vector-multiplication operation.
14. An analog device for effecting a linear transformation of input data corresponding to respective discrete values of a function, comprising:
- a plurality of input nodes for receiving analog signals representative of said input data;
- a plurality of analog phase-shift components connected to each of said plurality of input nodes for producing corresponding phase shifts in said analog signals; and
- a plurality of output nodes connected to said phase-shift components for producing analog output signals representative of output data corresponding to a linear transform of said discrete values of the function;
- wherein each of said phase-shift components includes at least one phase-shift element of a serially connected multiplicity of phase-shift elements, wherein said multiplicity of phase-shift elements produces a total phase shift equal to 2 kπ radians, k being an integer.
15. The device of claim 14, wherein each of said phase-shift elements produces an equal phase shift.
16. An identity transform oscillator comprising:
- a plurality of input nodes for receiving analog signals representative of input data;
- a plurality of analog phase-shift components connected to each of said plurality of input nodes for producing corresponding phase shifts in said analog signals;
- a plurality of output nodes connected to said phase-shift components for producing analog output signals representative of output data corresponding to a linear transform of said input data;
- a plurality of inverse-transform input nodes for receiving said analog output signals;
- a plurality of analog inverse-transform phase-shift components connected to each of said plurality of inverse-transform input nodes for producing corresponding phase shifts in said analog output signals;
- a plurality of inverse-transform output nodes connected to said inverse-transform phase-shift components for producing analog inverse-transform output signals; and
- a feedback loop connecting each of said inverse-transform output nodes to a corresponding input node.
17. The device of claim 16, further comprising:
- a plurality of amplitude-control elements corresponding to said phase-shift components to adjust respective outputs thereof to desired transform amplitudes; and
- a plurality of inverse-transform amplitude-control elements corresponding to said inverse-transform phase-shift components to adjust respective inverse-transform outputs thereof to desired inverse-transform amplitudes.
18. The device of claim 16, wherein each of said phase-shift components includes a waveguide selected with a length suitable for producing a predetermined phase shift.
19. The device of claim 16, wherein each of said phase-shift components includes at least one phase-shift element of a serially connected multiplicity of phase-shift elements, wherein said multiplicity of phase-shift elements produces a total phase shift equal to 2 kπ radians, k being an integer.
20. The device of claim 16, wherein each of said inverse-transform phase-shift components includes a waveguide selected with a length suitable for producing a predetermined phase shift.
21. The device of claim 16, wherein each of said inverse-transform phase-shift components includes at least one phase-shift element of a serially connected multiplicity of phase-shift elements, wherein said multiplicity of phase-shift elements produces a total phase shift equal to 2 kπ radians, k being an integer.
22. A method for effecting a linear transformation of input data corresponding to respective discrete values of a function, comprising the steps of:
- feeding analog signals representative of said input data to a plurality of input nodes;
- phase shifting said analog signals using a plurality of analog phase-shift components connected to said plurality of input nodes so as to produce phase-shifted analog signals; and
- producing analog output signals representative of output data corresponding to a linear transform of said discrete values of the function at a plurality of output nodes connected to said phase-shift components;
- wherein said phase-shifting step is carried out using phase-shift components that include a waveguide selected with a length suitable for producing a predetermined phase shift.
23. The method of claim 22, further comprising the step of:
- adjusting an amplitude of each of said phase-shifted analog signals using a plurality of amplitude-control elements.
24. The method of claim 23, wherein said amplitude adjusting step is carried out using a programmable resistor in each of said amplitude-control elements.
25. The method of claim 22, further including the steps of performing an operation on said analog output signals, and of inverse transforming processed output data of said operation.
26. The method of claim 25, wherein said step of inverse transforming said processed output data of the operation is carried out by:
- phase shifting said processed output data using a plurality of analog inverse-transform phase-shift components to produce inverse-transform phase-shifted analog signals; and
- utilizing said inverse-transform phase-shifted analog signals after amplitude adjustment to produce analog inverse-transform output signals.
27. The method of claim 26, further comprising the step of:
- adjusting an amplitude of each of said phase-shifted analog signals using a plurality of amplitude-control elements; and
- adjusting an amplitude of each of said inverse-transform phase-shifted analog signals using a plurality of inverse-transform amplitude-control elements.
28. A method for effecting a linear transformation of input data corresponding to respective discrete values of a function, comprising the steps of:
- feeding analog signals representative of said input data to a plurality of input nodes;
- phase shifting said analog signals using a plurality of analog phase-shift components connected to said plurality of input nodes so as to produce phase-shifted analog signals; and
- producing analog output signals representative of output data corresponding to a linear transform of said discrete values of the function at a plurality of output nodes connected to said phase-shift components;
- wherein said phase-shifting step is carried out using phase-shift components that include at least one phase-shift element of a serially connected multiplicity of phase-shift elements, wherein said multiplicity of phase-shift elements produces a total phase shift equal to 2 kπ radians, k being an integer.
29. The method of claim 28, wherein each of said phase-shift elements produces an equal phase shift.
5237629 | August 17, 1993 | Hietala et al. |
5293406 | March 8, 1994 | Suzuki |
5440270 | August 8, 1995 | Sevastopoulos et al. |
5652772 | July 29, 1997 | Isaksson et al. |
5959875 | September 28, 1999 | Kawahara et al. |
5987005 | November 16, 1999 | Fertner et al. |
6175671 | January 16, 2001 | Roberts |
6211734 | April 3, 2001 | Ahn |
6271690 | August 7, 2001 | Hirano et al. |
6295547 | September 25, 2001 | Zhang et al. |
6411417 | June 25, 2002 | Roberts et al. |
Type: Grant
Filed: Jan 21, 2003
Date of Patent: Oct 11, 2005
Patent Publication Number: 20040141459
Inventor: Frank A. Tinker (Tucson, AZ)
Primary Examiner: Andy Lee
Attorney: Quarles & Brady Streich Lang LLP
Application Number: 10/348,236